[SPARK-16827] Stop reporting spill metrics as shuffle metrics

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.0.0
• Component/s: Shuffle, Spark Core
• Labels: performance
• Target Version/s: 2.0.3, 2.1.0
• Assignee: Brian Cho
• Reporter: Sital Kedia
• Created: 01/Aug/16 00:24
• Updated: 12/Nov/16 17:27
• Resolved: 12/Oct/16 23:01

Description

One of our hive job which looks like this -

 SELECT  userid
     FROM  table1 a
     JOIN table2 b
      ON    a.ds = '2016-07-15'
      AND  b.ds = '2016-07-15'
      AND  a.source_id = b.id

After upgrade to Spark 2.0 the job is significantly slow. Digging a little into it, we found out that one of the stages produces excessive amount of shuffle data. Please note that this is a regression from Spark 1.6. Stage 2 of the job which used to produce 32KB shuffle data with 1.6, now produces more than 400GB with Spark 2.0. We also tried turning off whole stage code generation but that did not help.

PS - Even if the intermediate shuffle data size is huge, the job still produces accurate output.

Issue Links

• links to: [Github] Pull Request #15347 (dafrista)
• links to: [Github] Pull Request #15455 (dafrista)
• links to: [Github] Pull Request #15616 (dreamworks007)